From patchwork Wed Dec 19 15:55:47 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Joseph Salisbury X-Patchwork-Id: 1016153 Return-Path: X-Original-To: incoming@patchwork.ozlabs.org Delivered-To: patchwork-incoming@bilbo.ozlabs.org Authentication-Results: ozlabs.org; spf=none (mailfrom) smtp.mailfrom=lists.ubuntu.com (client-ip=91.189.94.19; helo=huckleberry.canonical.com; envelope-from=kernel-team-bounces@lists.ubuntu.com; receiver=) Authentication-Results: ozlabs.org; dmarc=fail (p=none dis=none) header.from=canonical.com Received: from huckleberry.canonical.com (huckleberry.canonical.com [91.189.94.19]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ozlabs.org (Postfix) with ESMTPS id 43Kfct4f9qz9sBh; Thu, 20 Dec 2018 02:55:58 +1100 (AEDT) Received: from localhost ([127.0.0.1] helo=huckleberry.canonical.com) by huckleberry.canonical.com with esmtp (Exim 4.86_2) (envelope-from ) id 1gZeCP-0003DU-9l; Wed, 19 Dec 2018 15:55:53 +0000 Received: from youngberry.canonical.com ([91.189.89.112]) by huckleberry.canonical.com with esmtps (TLS1.0:DHE_RSA_AES_128_CBC_SHA1:128) (Exim 4.86_2) (envelope-from ) id 1gZeCL-0003BH-0i for kernel-team@lists.ubuntu.com; Wed, 19 Dec 2018 15:55:49 +0000 Received: from 1.general.jsalisbury.us.vpn ([10.172.67.212] helo=salisbury) by youngberry.canonical.com with esmtpsa (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.76) (envelope-from ) id 1gZeCK-00064T-K5 for kernel-team@lists.ubuntu.com; Wed, 19 Dec 2018 15:55:48 +0000 Received: by salisbury (Postfix, from userid 1000) id 9CF967E1AEB; Wed, 19 Dec 2018 10:55:47 -0500 (EST) From: Joseph Salisbury To: kernel-team@lists.ubuntu.com Subject: [SRU][Cosmic][PATCH 2/2] Drivers: hv: vmbus: Offload the handling of channels to two workqueues Date: Wed, 19 Dec 2018 10:55:47 -0500 Message-Id: X-Mailer: git-send-email 2.17.1 In-Reply-To: References: In-Reply-To: References: X-BeenThere: kernel-team@lists.ubuntu.com X-Mailman-Version: 2.1.20 Precedence: list List-Id: Kernel team discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , MIME-Version: 1.0 Errors-To: kernel-team-bounces@lists.ubuntu.com Sender: "kernel-team" From: Dexuan Cui BugLink: https://bugs.launchpad.net/bugs/1807757 vmbus_process_offer() mustn't call channel->sc_creation_callback() directly for sub-channels, because sc_creation_callback() -> vmbus_open() may never get the host's response to the OPEN_CHANNEL message (the host may rescind a channel at any time, e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind() may not wake up the vmbus_open() as it's blocked due to a non-zero vmbus_connection.offer_in_progress, and finally we have a deadlock. The above is also true for primary channels, if the related device drivers use sync probing mode by default. And, usually the handling of primary channels and sub-channels can depend on each other, so we should offload them to different workqueues to avoid possible deadlock, e.g. in sync-probing mode, NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() -> rtnl_lock(), and causes deadlock: the former gets the rtnl_lock and waits for all the sub-channels to appear, but the latter can't get the rtnl_lock and this blocks the handling of sub-channels. The patch can fix the multiple-NIC deadlock described above for v3.x kernels (e.g. RHEL 7.x) which don't support async-probing of devices, and v4.4, v4.9, v4.14 and v4.18 which support async-probing but don't enable async-probing for Hyper-V drivers (yet). The patch can also fix the hang issue in sub-channel's handling described above for all versions of kernels, including v4.19 and v4.20-rc4. So actually the patch should be applied to all the existing kernels, not only the kernels that have 8195b1396ec8. Fixes: 8195b1396ec8 ("hv_netvsc: fix deadlock on hotplug") Cc: stable@vger.kernel.org Cc: Stephen Hemminger Cc: K. Y. Srinivasan Cc: Haiyang Zhang Signed-off-by: Dexuan Cui Signed-off-by: K. Y. Srinivasan Signed-off-by: Greg Kroah-Hartman (backported from commit 37c2578c0c40e286bc0d30bdc05290b2058cf66e) Signed-off-by: Joseph Salisbury --- drivers/hv/channel_mgmt.c | 190 +++++++++++++++++++++++++++++++--------------- drivers/hv/connection.c | 24 +++++- drivers/hv/hyperv_vmbus.h | 7 ++ include/linux/hyperv.h | 7 ++ 4 files changed, 162 insertions(+), 66 deletions(-) diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c index 0f0e091..70720ba 100644 --- a/drivers/hv/channel_mgmt.c +++ b/drivers/hv/channel_mgmt.c @@ -447,61 +447,16 @@ void vmbus_free_channels(void) } } -/* - * vmbus_process_offer - Process the offer by creating a channel/device - * associated with this offer - */ -static void vmbus_process_offer(struct vmbus_channel *newchannel) +/* Note: the function can run concurrently for primary/sub channels. */ +static void vmbus_add_channel_work(struct work_struct *work) { - struct vmbus_channel *channel; - bool fnew = true; + struct vmbus_channel *newchannel = + container_of(work, struct vmbus_channel, add_channel_work); + struct vmbus_channel *primary_channel = newchannel->primary_channel; unsigned long flags; u16 dev_type; int ret; - /* Make sure this is a new offer */ - mutex_lock(&vmbus_connection.channel_mutex); - - /* - * Now that we have acquired the channel_mutex, - * we can release the potentially racing rescind thread. - */ - atomic_dec(&vmbus_connection.offer_in_progress); - - list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) { - if (!uuid_le_cmp(channel->offermsg.offer.if_type, - newchannel->offermsg.offer.if_type) && - !uuid_le_cmp(channel->offermsg.offer.if_instance, - newchannel->offermsg.offer.if_instance)) { - fnew = false; - break; - } - } - - if (fnew) - list_add_tail(&newchannel->listentry, - &vmbus_connection.chn_list); - - mutex_unlock(&vmbus_connection.channel_mutex); - - if (!fnew) { - /* - * Check to see if this is a sub-channel. - */ - if (newchannel->offermsg.offer.sub_channel_index != 0) { - /* - * Process the sub-channel. - */ - newchannel->primary_channel = channel; - spin_lock_irqsave(&channel->lock, flags); - list_add_tail(&newchannel->sc_list, &channel->sc_list); - channel->num_sc++; - spin_unlock_irqrestore(&channel->lock, flags); - } else { - goto err_free_chan; - } - } - dev_type = hv_get_dev_type(newchannel); init_vp_index(newchannel, dev_type); @@ -519,27 +474,26 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel) /* * This state is used to indicate a successful open * so that when we do close the channel normally, we - * can cleanup properly + * can cleanup properly. */ newchannel->state = CHANNEL_OPEN_STATE; - if (!fnew) { - struct hv_device *dev - = newchannel->primary_channel->device_obj; + if (primary_channel != NULL) { + /* newchannel is a sub-channel. */ + struct hv_device *dev = primary_channel->device_obj; if (vmbus_add_channel_kobj(dev, newchannel)) - goto err_free_chan; + goto err_deq_chan; + + if (primary_channel->sc_creation_callback != NULL) + primary_channel->sc_creation_callback(newchannel); - if (channel->sc_creation_callback != NULL) - channel->sc_creation_callback(newchannel); newchannel->probe_done = true; return; } /* - * Start the process of binding this offer to the driver - * We need to set the DeviceObject field before calling - * vmbus_child_dev_add() + * Start the process of binding the primary channel to the driver */ newchannel->device_obj = vmbus_device_create( &newchannel->offermsg.offer.if_type, @@ -568,13 +522,28 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel) err_deq_chan: mutex_lock(&vmbus_connection.channel_mutex); - list_del(&newchannel->listentry); + + /* + * We need to set the flag, otherwise + * vmbus_onoffer_rescind() can be blocked. + */ + newchannel->probe_done = true; + + if (primary_channel == NULL) { + list_del(&newchannel->listentry); + } else { + spin_lock_irqsave(&primary_channel->lock, flags); + list_del(&newchannel->sc_list); + spin_unlock_irqrestore(&primary_channel->lock, flags); + } + mutex_unlock(&vmbus_connection.channel_mutex); if (newchannel->target_cpu != get_cpu()) { put_cpu(); smp_call_function_single(newchannel->target_cpu, - percpu_channel_deq, newchannel, true); + percpu_channel_deq, + newchannel, true); } else { percpu_channel_deq(newchannel); put_cpu(); @@ -582,14 +551,104 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel) vmbus_release_relid(newchannel->offermsg.child_relid); -err_free_chan: free_channel(newchannel); } /* + * vmbus_process_offer - Process the offer by creating a channel/device + * associated with this offer + */ +static void vmbus_process_offer(struct vmbus_channel *newchannel) +{ + struct vmbus_channel *channel; + struct workqueue_struct *wq; + unsigned long flags; + bool fnew = true; + + mutex_lock(&vmbus_connection.channel_mutex); + + /* + * Now that we have acquired the channel_mutex, + * we can release the potentially racing rescind thread. + */ + atomic_dec(&vmbus_connection.offer_in_progress); + + list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) { + if (!uuid_le_cmp(channel->offermsg.offer.if_type, + newchannel->offermsg.offer.if_type) && + !uuid_le_cmp(channel->offermsg.offer.if_instance, + newchannel->offermsg.offer.if_instance)) { + fnew = false; + break; + } + } + + if (fnew) + list_add_tail(&newchannel->listentry, + &vmbus_connection.chn_list); + else { + /* + * Check to see if this is a valid sub-channel. + */ + if (newchannel->offermsg.offer.sub_channel_index == 0) { + mutex_unlock(&vmbus_connection.channel_mutex); + /* + * Don't call free_channel(), because newchannel->kobj + * is not initialized yet. + */ + kfree(newchannel); + WARN_ON_ONCE(1); + return; + } + /* + * Process the sub-channel. + */ + newchannel->primary_channel = channel; + spin_lock_irqsave(&channel->lock, flags); + list_add_tail(&newchannel->sc_list, &channel->sc_list); + spin_unlock_irqrestore(&channel->lock, flags); + } + + mutex_unlock(&vmbus_connection.channel_mutex); + + /* + * vmbus_process_offer() mustn't call channel->sc_creation_callback() + * directly for sub-channels, because sc_creation_callback() -> + * vmbus_open() may never get the host's response to the + * OPEN_CHANNEL message (the host may rescind a channel at any time, + * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind() + * may not wake up the vmbus_open() as it's blocked due to a non-zero + * vmbus_connection.offer_in_progress, and finally we have a deadlock. + * + * The above is also true for primary channels, if the related device + * drivers use sync probing mode by default. + * + * And, usually the handling of primary channels and sub-channels can + * depend on each other, so we should offload them to different + * workqueues to avoid possible deadlock, e.g. in sync-probing mode, + * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() -> + * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock + * and waits for all the sub-channels to appear, but the latter + * can't get the rtnl_lock and this blocks the handling of + * sub-channels. + */ + INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work); + wq = fnew ? vmbus_connection.handle_primary_chan_wq : + vmbus_connection.handle_sub_chan_wq; + queue_work(wq, &newchannel->add_channel_work); +} + +/* * We use this state to statically distribute the channel interrupt load. */ static int next_numa_node_id; +/* + * init_vp_index() accesses global variables like next_numa_node_id, and + * it can run concurrently for primary channels and sub-channels: see + * vmbus_process_offer(), so we need the lock to protect the global + * variables. + */ +static DEFINE_SPINLOCK(bind_channel_to_cpu_lock); /* * Starting with Win8, we can statically distribute the incoming @@ -623,6 +682,8 @@ static void init_vp_index(struct vmbus_channel *channel, u16 dev_type) return; } + spin_lock(&bind_channel_to_cpu_lock); + /* * Based on the channel affinity policy, we will assign the NUMA * nodes. @@ -704,6 +765,9 @@ static void init_vp_index(struct vmbus_channel *channel, u16 dev_type) channel->target_cpu = cur_cpu; channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu); + + spin_unlock(&bind_channel_to_cpu_lock); + } static void vmbus_wait_for_unload(void) diff --git a/drivers/hv/connection.c b/drivers/hv/connection.c index f4d08c8..4fe117b7 100644 --- a/drivers/hv/connection.c +++ b/drivers/hv/connection.c @@ -190,6 +190,20 @@ int vmbus_connect(void) goto cleanup; } + vmbus_connection.handle_primary_chan_wq = + create_workqueue("hv_pri_chan"); + if (!vmbus_connection.handle_primary_chan_wq) { + ret = -ENOMEM; + goto cleanup; + } + + vmbus_connection.handle_sub_chan_wq = + create_workqueue("hv_sub_chan"); + if (!vmbus_connection.handle_sub_chan_wq) { + ret = -ENOMEM; + goto cleanup; + } + INIT_LIST_HEAD(&vmbus_connection.chn_msg_list); spin_lock_init(&vmbus_connection.channelmsg_lock); @@ -280,10 +294,14 @@ void vmbus_disconnect(void) */ vmbus_initiate_unload(false); - if (vmbus_connection.work_queue) { - drain_workqueue(vmbus_connection.work_queue); + if (vmbus_connection.handle_sub_chan_wq) + destroy_workqueue(vmbus_connection.handle_sub_chan_wq); + + if (vmbus_connection.handle_primary_chan_wq) + destroy_workqueue(vmbus_connection.handle_primary_chan_wq); + + if (vmbus_connection.work_queue) destroy_workqueue(vmbus_connection.work_queue); - } if (vmbus_connection.int_page) { free_pages((unsigned long)vmbus_connection.int_page, 0); diff --git a/drivers/hv/hyperv_vmbus.h b/drivers/hv/hyperv_vmbus.h index 72eaba3..87d3d7d 100644 --- a/drivers/hv/hyperv_vmbus.h +++ b/drivers/hv/hyperv_vmbus.h @@ -335,7 +335,14 @@ struct vmbus_connection { struct list_head chn_list; struct mutex channel_mutex; + /* + * An offer message is handled first on the work_queue, and then + * is further handled on handle_primary_chan_wq or + * handle_sub_chan_wq. + */ struct workqueue_struct *work_queue; + struct workqueue_struct *handle_primary_chan_wq; + struct workqueue_struct *handle_sub_chan_wq; }; diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h index 5389012..56451ae 100644 --- a/include/linux/hyperv.h +++ b/include/linux/hyperv.h @@ -889,6 +889,13 @@ struct vmbus_channel { bool probe_done; + /* + * We must offload the handling of the primary/sub channels + * from the single-threaded vmbus_connection.work_queue to + * two different workqueue, otherwise we can block + * vmbus_connection.work_queue and hang: see vmbus_process_offer(). + */ + struct work_struct add_channel_work; }; static inline bool is_hvsock_channel(const struct vmbus_channel *c)